1. N o, you don’t understand, I mean… Irini Nomikou supervisor: Dr. Floriana Grasso The one with the conductor and the girl on the train Cond: Did you pay dear? Girl: No Cond: Where did you get on? Girl: Hindley Cond: But that's a couple of stations back. Have you got your card with you? Girl: Yeah Cond: It's 25.45. There is a surcharge, a fine, for avoiding paying your fare Girl: What??? But I...I just wasn't listening. Oh, please… Cond: Just kidding dear. £1.05 it is. Girl: Oh, all right. (Ha, ha, ha). Thanks! A computer that really understands what you say? Our work is in Natural Language Processing (NLP). NLP is a subfield of artificial intelligence and linguistics. It studies the problems of automated generation and understanding of natural human languages. In particular, our work is on misunderstandings. When in conversation, the hearer commits herself to an interpretation of what the speaker has said, in order to plan her response. Misunderstandings occur when the hearer's understanding is different to that intended by the speaker. Consider this example on the left. The conductor of the train attempts to joke with the girl but she obtains a different interpretation of his utterance. The conductor detects the misunderstanding and attempts to repair it. Now how would a computer be able to do that? But, how often do misunderstandings happen? Earlier views on misunderstandings suggested these are a deviation from the understanding norm. We adopt the recent view of misunderstandings as a normal phenomenon in communication. Complicated as it is, it provides the ground for observation and provides insight into the cognitive processes taking part in the process of interpreting as well as generating utterances. Misunderstandings take place in conversation way too often. Just observe people chatting around you, in the streets, on the bus, at university. Or even pay more attention when you chat with your friends. At least that is what we did and we discovered that yes, misunderstandings occur all the time. And we put our findings in a small collection of conversations (corpus) in order to analyse them and describe the phenomenon. The example you just read is part of that corpus. The analysis of the examples helped us understand the phenomenon and explain the process of detecting and repairing misunderstandings. What is exactly relevance theory? We instinctively know what relevant is, but how do we define that? Sperber and Wilson have developed Relevance Theory, a cognitive theory defining the phenomenon of relevance. The version we employ in our model: RelevantTo = believes(H, Update(INFERREDMEANING, CONTEXT)) ==> believes(H, relevant(P)) This is an example of a Prolog-like piece of code which tells us that if the INFERREDMEANING of the utterance (what the hearer believes the speaker has said) causes any update in the hearer’s CONTEXT (any change in the Hearer’s CONTEXT) then the Hearer would consequently believe that utterance (P) is relevant. In the absence of an Update the hearer (presupposing the Speaker’s utterance is always relevant - otherwise why uttering it) assumes there is a misunderstanding and attempts to reconstruct her context so that the INFERREDMEANING would cause an update to her CONTEXT, therefore detecting where the misunderstanding has taken place. In our train example we see that the conductor when interpreting the girl’s utterance gets no update on his CONTEXT – hence he detects a misunderstanding - so he tries to reformulate it and sees that the girl has actually taken it seriously and only when he temporarily adds that to his CONTEXT, he gets an update to it. Then he attempts to repair it by calling to the girl’s attention to modify her CONTEXT accordingly. The implementation Our model will be able to handle cases of misunderstandings in conversation. How will we do that? It will be incorporated in an AI language planner that can plan utterances and generate their semantics. The planner is written in Prolog, a programming language. It uses information about the speaker’s and hearer’s beliefs, goals and plans in order to plan an utterance. The implementation will be mainly used as an evaluation of the theory and the data for that evaluation will be the formal representations of the examples collected in our corpus. The computational model Our model combines a version of relevance theory and an account of context, beliefs, goals and plans. The context is a term often used in linguistics. The context is the sum of all the pragmatics information in a conversation (time, place, participants etc). Context = [Discourse, Hearer’sBeliefs, Hearer’sGoals, Hearer’sPlans, Hearer’sBeliefs on(Speaker’sBeliefs, Goals, Plans)]